AI model deprecation calendar: OpenAI, Anthropic, Google

Model retirements arrive one vendor page at a time, each with its own vocabulary and its own idea of what a date means. This page puts the announced ones in one place: every retirement from today, 13 October 2026, to mid-2027 that OpenAI, Anthropic and Google list on their own pages, plus the platforms that host those models on a different schedule. Each row names its source. Nothing here comes from a roundup.

It is reviewed monthly. The review re-reads every source, adds new announcements and removes dates that have passed. The date at the top of the page is the last time that happened.

For what to do once a date in this table applies to you, use the model deprecation migration playbook. For the 23 October OpenAI shutdown specifically, see OpenAI's October 23 shutdowns.

Read the dates the way each vendor means them

A date in one vendor's table is not the same kind of promise as a date in another's.

• OpenAI shuts a model down on the listed date: at that point it "will no longer be accessible"¹.
• Anthropic lists firm retirement dates for deprecated models, and "Not sooner than" floors for active ones. Requests to a retired model fail, and Anthropic gives at least 60 days' notice before retiring a publicly released model².
• Google's Gemini API lists the "earliest possible dates" a model might be shut down and promises notice of the exact date³.
• Vertex AI says its retirement timelines "may be extended" but "won't be moved to an earlier date than what is listed"⁴.
• Azure OpenAI returns 410 Gone after retirement, may auto-upgrade Standard deployments to the named replacement, and does not auto-upgrade provisioned deployments⁵.
• Amazon Bedrock puts the end-of-life date on each model card, says requests fail after it, and says migration "will not happen automatically"⁶.

The calendar

Dates are the vendor's own. "Listed replacement" is the replacement the vendor names; where a vendor names none, the cell says so.

Three things in the table deserve a second look.

Gemini 2.5 depends on where you call it. Vertex AI lists gemini-2.5-pro, gemini-2.5-flash and gemini-2.5-flash-lite for 20 October 2026⁴. The Gemini API's page says 2.5 models "are not deprecated and will continue to be served until further notice", limits access to users who have used them before, and recommends 3.5 Flash-Lite or 3.8 Flash for new projects³. The same model name, two different answers, one week apart.

OpenAI's December date is the next big one. The 11 December retirement takes the dated gpt-5, gpt-5-mini, gpt-5-nano and gpt-5-pro snapshots and o3 and o3-pro¹. Teams that just moved off o4-mini onto a GPT-5 snapshot should check which one.

OpenAI is retiring platform features, not only models. Evals, Agent Builder and the v1/prompts API end on 30 November¹. A prompt stored as a reusable prompt object has to move into your own code or configuration before then.

Not yet announced: Anthropic's earliest dates

Anthropic's model status table gives every active model a tentative retirement that is "Not sooner than" a date. These are floors, not announcements: Anthropic deprecates a model first, then gives at least 60 days' notice². Two of them fall this year.

Anthropic also deprecates parameters, not only models: on Claude 4.7 and later, a non-default temperature, top_p or top_k returns a 400, and its Python SDK from v1.0 removes them². When a Claude 4.5 caller moves to a 5.x replacement, check the sampling parameters it sends.

One price change: Gemini 3.8 Flash

The Gemini API pricing page lists Gemini 3.8 Flash at $0.75 input and $3.75 output per million tokens through 31 December 2026, and $1.50 and $7.50 from 1 January 2027; Gemini 3.6 Flash follows the same schedule, and the context caching, Batch, Flex and Priority prices for 3.8 Flash change on the same date⁸. It is not a retirement, but it changes the cost of every request on the same day for anyone who stays, which is why it belongs on a calendar of dates that need a decision. Budgets and cost forecasts that span the new year should use the 2027 rate.

Hosted copies keep their own schedules

Calling a model through a cloud platform means calling that platform's copy, on that platform's dates.

Azure OpenAI. Microsoft's schedule differs from OpenAI's in both directions⁹:

Microsoft's schedule also lists Claude models on Foundry: claude-sonnet-4-5 for 2026-11-30, and claude-haiku-4-5 for 2026-11-15 and claude-opus-4-5 for 2026-11-24⁹. Anthropic's own page, which says its dates apply to Microsoft Foundry as well, lists Haiku 4.5 and Opus 4.5 as active with no retirement announced². If you run Claude on Foundry, confirm with both before you plan around either.

Amazon Bedrock. Bedrock's lifecycle page says its dates "may differ from dates published by model providers" and that for Bedrock usage "only the dates on the model card apply"⁶. Read the model card for each Bedrock model id you call.

Google Cloud. Anthropic says Bedrock and Google Cloud set their own retirement schedules for Claude², and Vertex AI's own Gemini dates differ from the Gemini API's, as the 2.5 rows show.

Vendors with no dates in range

Mistral publishes a deprecated and retired models table on its models overview; the only retirement after today is zai-glm-5-2 on 31 October 2026⁷. Cohere's deprecations page lists nothing retiring after 13 October 2026; its latest entry took effect on 4 April 2026¹⁰. xAI's models page publishes no retirement dates; it documents aliases such as <modelname>-latest that move to newer versions¹¹, which means the model behind an alias can change without a retirement notice. Azure's schedule lists Grok retirements for its own deployments⁹.

Using the calendar

A calendar only helps if each row is checked against what you actually send. For every row:

• Search your traffic for the exact id, and the ids it aliases, over at least 30 days.
• Note which platform each caller uses — the vendor's API, Azure, Bedrock, Vertex AI or an aggregator — and take that platform's date.
• Qualify the listed replacement on recorded requests, with the parameters your code sends.
• Put the model behind a name you control before the cut-over, not during it.
• Refuse the old id in your own systems a few days before the vendor does.

Tracking these dates with INFRO

INFRO is a control plane for the provider accounts you already hold: you connect your own OpenAI, Anthropic, Google, Azure, Bedrock or Mistral account, the provider bills you directly, and INFRO records each request that passes through it. For a calendar like this one, it answers "does this row apply to us?" from traffic, and turns the answer into configuration.

• Who calls a model. GET /v1/usage groups requests and estimated provider cost by model, provider, connection, project and key; GET /v1/requests filters by the same fields. History reaches back 7, 14, 30 or 90 days on Developer, Startup, Team and Business. See usage analytics.
• Stop new usage. A model allowlist entry such as !openai/gpt-image-1.5 blocks one model at organization, project or key level; requests for it are refused with 403 model_not_allowed before any provider call.
• Move without a deploy. Applications that name a route instead of a model id change model when the route's targets change.
• Notices. When INFRO deprecates a model in its own catalog, owners and admins get an email and a notice under the console's bell with the retirement date, the successor, and the requests, keys and routes in the organization that still use it. After a cut-over, traffic changes flag a connection answered by a model id it had not reported in the previous 14 days.

INFRO's catalog has its own deprecation policy — at least 30 days' notice, after which a retired catalog id becomes an alias for its designated successor — but that is a catalog convenience, not a stay of execution. The provider whose account you connect decides whether a model answers on its date.